Issues

Does the Dolma 2 pre-tokenizer fix change hero-small SFT at the full anchor budget? (ledger H016, aligned-11768)

#1912

· abhishekraok opened

on Sep 21, 2026

main violates its own ban-keywords pre-commit hook, and CI never runs that hook

#1906

· VishnuR23 opened

on Sep 20, 2026

Non-termination on the Dolci-Think SFT anchors: SFT-budget and proxy-vs-hero 2x2 on matched Math-500 (ledger H012/H013)

#1903

· abhishekraok opened

on Sep 19, 2026

Olmo Identity Data

#1902

· mgmorgan23 opened

on Sep 18, 2026

Mismatch between allenai/Olmo-3-7B-RL-Zero-Code chat_template.jinja and the training template displayed in commit 70b0472af73bd6428e14c5d73abebbabaaeb4997

#1899

· 1BIMU opened

on Sep 17, 2026

Dolci-Think SFT anchor on the hero small PT-EMO-only checkpoint: paired arm for the EMO-in-pretraining question

#1898

· abhishekraok opened

on Sep 16, 2026

transformers 5 loads Dolma 2 tokenizer-only repos with the GPT-2 pre-tokenizer, not the cl100k regex in tokenizer.json: SFT segmentation differs from pretraining (+9.8% tokens, 72% of Dolci-Think messages)

#1896

· abhishekraok opened

on Sep 14, 2026

Dolci-Think SFT anchor on the multi-trillion-token KDA checkpoint: re-establish the #1853 anchor and measure the pretraining-token effect

#1895

· abhishekraok opened

on Sep 14, 2026

Verify evals on Papers with Code

#1894

· NielsRogge opened

on Sep 14, 2026

Fix terminal evaluation recovery and SFT data filtering for the KDA checkpoint

#1891

· abhishekraok opened

on Sep 11, 2026

Hub think-dev tokenizer breaks SFT label derivation; #1879 does not reach it

#1882

· abhishekraok opened

on Sep 3, 2026

Anneal-matched 0.5-epoch anchor for the Olmo 3.5 SFT mixture search (KDA 1.2B, Dolci-Think, seq 65536)

#1880

· abhishekraok opened

on Sep 3, 2026

Navigation Menu

All issues

Does the Dolma 2 pre-tokenizer fix change hero-small SFT at the full anchor budget? (ledger H016, aligned-11768)

main violates its own ban-keywords pre-commit hook, and CI never runs that hook

Non-termination on the Dolci-Think SFT anchors: SFT-budget and proxy-vs-hero 2x2 on matched Math-500 (ledger H012/H013)

Olmo Identity Data

Mismatch between allenai/Olmo-3-7B-RL-Zero-Code chat_template.jinja and the training template displayed in commit 70b0472af73bd6428e14c5d73abebbabaaeb4997

Dolci-Think SFT anchor on the hero small PT-EMO-only checkpoint: paired arm for the EMO-in-pretraining question

transformers 5 loads Dolma 2 tokenizer-only repos with the GPT-2 pre-tokenizer, not the cl100k regex in tokenizer.json: SFT segmentation differs from pretraining (+9.8% tokens, 72% of Dolci-Think messages)

Dolci-Think SFT anchor on the multi-trillion-token KDA checkpoint: re-establish the #1853 anchor and measure the pretraining-token effect

Verify evals on Papers with Code

Fix terminal evaluation recovery and SFT data filtering for the KDA checkpoint

Hub think-dev tokenizer breaks SFT label derivation; #1879 does not reach it

Anneal-matched 0.5-epoch anchor for the Olmo 3.5 SFT mixture search (KDA 1.2B, Dolci-Think, seq 65536)

Search results